Modelpublic

Abhay2310/qwen_vlm_0.8B_finetune_math_ocr

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes56downloads

Model Card

Qwen3.5-0.8B VLM — Equation Image → LaTeX

A fine-tune of unsloth/Qwen3.5-0.8B that transcribes images of mathematical equations into LaTeX. Quantized to Q4KM for CPU inference.

Files

File	Size	Purpose
`Qwen3.5-0.8B.Q4_K_M.gguf`	542 MB	Quantized language model
`Qwen3.5-0.8B.F16-mmproj.gguf`	205 MB	Vision projector — required

Both files are needed. Loading the model without --mmproj gives you a text-only model with no ability to see images.

Usage

With llama.cpp:

bash

llama-mtmd-cli \
  -m Qwen3.5-0.8B.Q4_K_M.gguf \
  --mmproj Qwen3.5-0.8B.F16-mmproj.gguf \
  --image equation.png \
  -p "Convert the equation in the image to LaTeX." \
  --temp 0.1

Low temperature is recommended — transcription is a deterministic task, and sampling introduces errors in brace matching and command names.

Training

Fine-tuned with Unsloth using LoRA, then merged to 16-bit and converted to GGUF.

Limitations

—0.8B parameters at 4-bit is a small, heavily compressed model. Expect errors on dense or unusual notation.
—Best results on clean screenshots; handwriting and low-resolution photos are harder.
—Output may need brace-matching cleanup before it compiles.